Cloudflare Uses an AI Harness to Probe and Harden Its WAF
InfoQ, Wednesday, October 7th, 2026
Cloudflare used frontier AI models in a testing harness to mutate blocked attacks, leading to three new WAF managed rule changes.
Cloudflare placed frontier AI models inside a controlled Python harness to probe its Web Application Firewall, starting from already blocked payloads and giving the models no access to rules or source code.
One model proposed mutations while another reviewed responses, and across 45 scenarios, 1,107 attempts produced 49 findings after human triage, nearly all command injection or SSRF.
The work led to two new SSRF detections and an improved SSRF cloud rule in Cloudflare's Managed Ruleset, with humans serving as the final check.